ChatGPT Custom Instructions: How to Set Them Up Right (2026)

1,500 characters is all the space ChatGPT gives you in each custom instructions box, and most people waste them. They write a short autobiography in one and a polite wish list in the other, then wonder why answers still feel generic. Used well, the same two boxes are the single most useful setting in the whole app: a standing brief that quietly improves every chat without you typing a word of setup again. The difference is entirely in how you fill them.

This is a precise, current guide to doing that, for 2026. Where the feature actually lives now, what belongs in each field, copy-paste templates you can adapt, and the honest reason instructions sometimes get ignored, which almost no guide explains. Custom instructions work on the free plan as well as paid, so everything here applies whatever you're paying.

What custom instructions are, and where to find them

Custom instructions are a pair of saved settings that tell ChatGPT who you are and how to answer, applied automatically to every new chat so you stop repeating yourself. They're a standing brief, not a one-off prompt.

To find them, open Settings and choose Customize ChatGPT. On the web, click your profile in the bottom left first; on mobile, look under Settings, then Personalization. Make sure Enable customization is toggled on, and two fields appear:

• What would you like ChatGPT to know about you? Context about you and your work.
• How would you like ChatGPT to respond? Your rules for the output.

Each field holds about 1,500 characters. Alongside them sit a few related controls worth knowing: a base style and tone preset (Default, Professional, Friendly, Candid, and others), and characteristics that fine-tune things like brevity and emoji use. These stack with your written instructions rather than replacing them. One thing to note: saving applies to new chats going forward, not to conversations already open.

What should you write in each box?

Here's the mental model that fixes most weak setups. The first box is for context that changes how an answer should be written. The second is for rules ChatGPT should follow every time. People get this wrong by treating box one as a biography and box two as a politeness request.

Box one isn't a CV. "I'm a designer" is a fact the model rarely acts on. What helps is the context that shapes a useful reply: what you use ChatGPT for, who your work is aimed at, and how much you already know. Box two isn't where you ask it to be nice; it's where you set hard rules: how long answers should be, what format you want, the tone, and what to do when a request is unclear.

Two rules make the second box land. Write constraints, not descriptions, so "answer in plain English at around 200 words" beats "be clear and helpful." And whenever you tell ChatGPT to avoid something, say what to do instead, because a "don't" with no replacement leaves it guessing. Watch out for contradictions too: asking for replies that are both detailed and concise, or formal and casual, forces the model to pick one and ignore the other.

Compare a vague second box:

Be helpful and professional, and give me clear, high-quality answers.

With a sharp one that actually constrains the output:

Default to plain English at around 150 to 250 words. Lead with the answer, then the reasoning. Use bullet points for any list of three or more. Skip introductions and filler. If my request is ambiguous, ask one clarifying question before answering.

The first changes almost nothing. The second changes every reply you get.

Copy-paste templates by role

Use these as starting points and swap in your own details. Each shows both boxes; paste the first part into the "about you" field and the second into the "how to respond" field, on ChatGPT GPT-5.5 as of June 2026.

A solid general setup for most people:

ABOUT YOU: I use ChatGPT for work and everyday tasks across writing, planning and research. I'm comfortable with detail but short on time. HOW TO RESPOND: Lead with the answer. Keep it concise, plain English, no filler or flattery. Use bullets for lists. When facts matter, say how confident you are. If a request is unclear, ask before assuming.

If you mostly write business communication:

ABOUT YOU: I'm a manager who writes emails, briefs and updates for non-technical colleagues and clients. HOW TO RESPOND: Default to a clear, professional but human tone. Keep emails under 150 words. Give me the direct version first, then a softer alternative if the topic is sensitive. Avoid corporate jargon and exclamation marks.

Writers and content creators:

ABOUT YOU: I write articles and social posts in a conversational, concrete style and I edit heavily. HOW TO RESPOND: Draft in plain, specific language with short sentences. No clichés, no hype words, no em dashes. Offer two or three angle options before writing long. Treat your draft as a first pass I will rewrite, not a finished piece.

For a developer:

ABOUT YOU: I code mainly in Python and JavaScript and value readable, maintainable solutions. HOW TO RESPOND: Give the working code first, then a short explanation. Comment only the non-obvious parts. Flag edge cases and security concerns. If there's a simpler standard-library approach, show that before reaching for a dependency.

Students who want to learn, not just collect answers:

ABOUT YOU: I'm a student trying to understand topics, not just get answers handed to me. HOW TO RESPOND: Explain in plain terms with a simple example. For problems, walk me through the steps rather than only giving the result, and check my understanding with a question at the end. Keep it encouraging.

One widely shared template, maintained by developer Denis Shiryaev on GitHub, pushes this further by telling ChatGPT to adopt a relevant expert role and quietly grade its own draft against a rubric before replying. It's a neat trick, though for most people the plainer setups above are easier to live with and edit.

Why does ChatGPT ignore your custom instructions?

This is the part the template lists skip. Your instructions don't sit in some protected rulebook. They're turned into a short brief at the very start of each chat, then have to compete for attention with every message you send afterwards. In a long conversation, or after you upload a big file, that brief drifts toward the back of the model's attention and gets weighted less. They also cost you a little: instructions are sent with every chat and count against the model's working context, which is another reason to keep them lean.

On top of that, ChatGPT is trained to be agreeable and "helpful," so it leans toward adding context or reformatting even when you asked it not to, and it tends to drop negative instructions in favour of doing what feels complete.

So treat custom instructions as a default working style, not an unbreakable law. The fixes are practical: keep them short and filter-like, since the people who get the most from this feature trim what they don't need rather than pile on more; avoid contradictions; and when a long chat stops following them, start a fresh one and the brief is back at full strength. For one-off needs, repeat the key rule in the chat itself, which our guide to writing prompts covers in more depth.

Custom instructions versus Memory

These two get confused constantly, and the difference is simple. Custom instructions are static and yours to control: you write them, you edit them, and they apply the same way to every new chat. Memory is automatic: ChatGPT decides what to remember from your conversations and brings it back later, without you managing a list.

Use custom instructions for the stable stuff, your role, tone, and formatting rules. Let Memory handle the evolving details it picks up as you go. They're both part of ChatGPT's personalisation, and they work together, but only custom instructions give you an exact, editable brief you can read in full. For how these fit a bigger setup with Projects, our guide to building a second brain maps each feature to its job.

Can you have different custom instructions for different tasks?

You only get one global set, so you can't keep a separate "work" and "personal" version in the main settings. There are two clean ways around it. Projects each carry their own instructions that apply inside that workspace and override your global ones, which is ideal for keeping a client's context separate. A Custom GPT also has its own instructions built for one job, and you can switch to it when you need that mode; here's how to create a custom GPT from scratch. Both let you run more than one standing brief without rewriting your settings every time you change hats. These are the same boxes a system prompts setup uses, in different places.

How to turn custom instructions off

If you want a clean slate, open Settings, go to Customize ChatGPT, and toggle Enable customization off; your saved text stays but stops applying. For a single chat where you don't want them active, start a Temporary Chat instead, which ignores both custom instructions and Memory and won't be saved to your history. That's handy when you're testing how a raw prompt performs, or asking something unrelated to your usual work.

Keep it private

Because your instructions are sent on every chat, treat them as context, not a vault. Keep passwords, financial details, client secrets, and other people's personal information out of the boxes entirely. On the consumer plans, ChatGPT may use your inputs to improve its models unless you opt out, so if that matters, open Settings, then Data Controls, and switch model training off. None of this makes the feature risky; it just means a standing brief deserves the same care as anything else you type into a chatbot.

What this post does not cover

This is a guide to the custom instructions feature inside the ChatGPT app, not the developer system prompts you write through the API, which our ChatGPT prompt tips and system prompt guide touch on. Feature names, fields, and limits change often, so check OpenAI's own settings screen for the current layout before relying on a detail. The templates here are starting points to adapt, not finished assistants, and ChatGPT can still get things wrong, so keep checking its output rather than trusting the setup to make it perfect.

Sources

1. OpenAI Help Center: ChatGPT custom instructions
2. OpenAI Help Center: is memory different from custom instructions?
3. OpenAI Help Center: customizing your ChatGPT personality
4. OpenAI: custom instructions for ChatGPT